Motivation
Traditional inflation models struggle with qualitative data like Fed statements. I tested whether transformer models could outperform ARIMA by analyzing:
- FOMC meeting minutes (2008-2024)
- ECB press conference transcripts
- BIS financial stability reports
Technical Approach
Fine-tuned distilbert-base-uncased to classify text segments into:
# Sample PyTorch code
class InflationClassifier(nn.Module):
def __init__(self):
super().__init__()
self.bert = DistilBertModel.from_pretrained('distilbert-base-uncased')
self.classifier = nn.Linear(768, 3) # Hawkish/Dovish/Neutral
Accuracy: 82% (vs. 67% for sentiment analysis baselines)
Key Findings
Policy Tone Matters
Models using "dovish" language segments reduced prediction errors by 19%.
| Model | MAE | RMSE |
|---|